In the last post, we built the Wiener Process from scratch. We started with a coin flip, constructed a Binomial Tree, and took the continuous-time limit to obtain a cloud of uncertainty whose variance grows in lock-step with time. It is a beautiful mathematical object.

But beauty alone does not pay bills. A cloud of paths tells us what is possible, not what things are worth. If stock prices follow a Wiener Process, how do you price an option on that stock? How do you hedge it? How do you even write an equation for a price that depends on something random?

For that, we need a new kind of calculus. Welcome to Stochastic Calculus.

Why Regular Calculus Breaks

Recall from the last post that the Wiener Process $W(t)$ has the property:

W(t) - W(s) \sim \mathcal{N}(0,\ t - s)

The variance of any move equals the time elapsed. This sounds harmless, but it has a brutal consequence for calculus. In ordinary calculus, the chain rule lets you differentiate a composition of functions. If $f$ is a function of $x$ and $x$ is a function of $t$ , then:

\frac{df}{dt} = \frac{df}{dx} \cdot \frac{dx}{dt}

This works because in ordinary calculus, $(\Delta x)^2$ is so small relative to $\Delta x$ that we can discard it. In the language of calculus, second-order terms “vanish.” But with Brownian motion, that assumption collapses entirely. The move $\Delta W$ over a tiny interval $\Delta t$ is of order $\sqrt{\Delta t}$ , which means:

(\Delta W)^2 \approx \Delta t

The square of the Wiener increment is of the same order as the time step itself. It does not vanish. This single fact breaks the chain rule and forces us to rebuild calculus from scratch. This correction term that ordinary calculus throws away turns out to be the central object of the entire theory.

The Stochastic Differential Equation (SDE)

Before we fix the chain rule, let us first write down the model we actually want to use for stock prices. The most widely used model in quantitative finance is Geometric Brownian Motion (GBM):

dS = \mu S \, dt + \sigma S \, dW_t

Let us unpack this piece by piece:

$S(t)$ is the stock price at time $t$ .
$\mu$ is the drift: the average rate of return. It represents the deterministic trend pulling the price upward over time.
$\sigma$ is the volatility: a measure of how large the random shocks are.
$dW_t$ is an infinitesimal increment of the Wiener Process — a draw from a Normal distribution with mean zero and variance $dt$ .

You can read this equation as: “In every tiny instant of time, the percentage change in the stock price is partly predictable (drift) and partly random (volatility times a Wiener shock).”

Notice the $S$ multiplying both terms on the right-hand side. This is what makes the model geometric: the shocks are proportional to the current price level. A $10 shock matters much more when a stock trades at $15 than when it trades at $500. This multiplicative structure also ensures that $S(t)$ can never go negative. It’s a useful property that simpler arithmetic models (where $dS = a \, dt + b \, dW$ ) do not have.

import numpy as np
import matplotlib.pyplot as plt
 
# Parameters
S0 = 100       # Initial stock price
mu = 0.08      # Annual drift (8% expected return)
sigma = 0.20   # Annual volatility (20%)
T = 1.0        # 1 year
steps = 252    # Trading days
dt = T / steps
n_paths = 5    # Simulate 5 paths to visualise
 
np.random.seed(0)
time_axis = np.linspace(0, T, steps + 1)
paths = np.zeros((n_paths, steps + 1))
paths[:, 0] = S0
 
for i in range(steps):
    dW = np.random.normal(0, np.sqrt(dt), n_paths)
    # GBM discretisation: dS = mu*S*dt + sigma*S*dW
    paths[:, i+1] = paths[:, i] * (1 + mu * dt + sigma * dW)
 
plt.figure(figsize=(11, 5))
for p in paths:
    plt.plot(time_axis, p, lw=1.2, alpha=0.8)
plt.axhline(S0, color='black', linestyle='--', alpha=0.4, label='Starting Price')
plt.title("Geometric Brownian Motion: 5 Simulated Stock Paths")
plt.xlabel("Time (years)")
plt.ylabel("Stock Price ($)")
plt.legend()
plt.grid(True, alpha=0.3)
plt.show()

Run this, and you will see five completely different futures for the same stock, all starting from $100. Some climb, some fall; none go below zero. That multiplicative noise is doing its job. Notice that the paths fan out over time, this is the “diffusion” from Post 9, now applied to a price level rather than a position on a number line.

The Itô Integral: Defining the Stochastic Integral

If we want to solve the SDE from above (to find $S(t)$ as an explicit function) we need to integrate both sides. The deterministic part $\int \mu S \, dt$ is perfectly fine; ordinary calculus handles that. But what about the random part?

\int_0^t \sigma S \, dW_s

How do you integrate with respect to a Wiener Process? Think back to how the ordinary Riemann integral is defined: slice the interval $[0, t]$ into tiny pieces, evaluate the function at the left endpoint of each piece, multiply by the width of that piece, and sum everything up.

We do exactly the same thing for the stochastic integral:

\int_0^t f(W_s)\, dW_s = \lim_{N \to \infty} \sum_{k=0}^{N-1} f(W_{t_k}) \cdot (W_{t_{k+1}} - W_{t_k})

The crucial detail is the left endpoint: $f(W_{t_k})$ is evaluated before we observe the Wiener increment $(W_{t_{k+1}} - W_{t_k})$ . This is not just a mathematical nicety; it is a financial necessity. You cannot know what the market will do in the next instant before it happens. Using the left endpoint means you are committing to a position before the price moves, which is exactly what a trader does. A strategy that could peek into the next increment would be cheating, and the mathematics enforces this honestly.

This specific construction is called the Itô Integral, and it is the foundation of stochastic calculus.

The Surprising Answer: A Simple Example

Let us test our intuition. In ordinary calculus, we know that $\int_0^t x \, dx = t^2/2$ . What is the stochastic equivalent?

\int_0^t W_s\, dW_s = ?

If stochastic calculus worked like regular calculus, the answer would simply be $\frac{1}{2}W_t^2$ . But it does not. Working through the Itô construction carefully, using the independence of Wiener increments and the fact that each increment has variance $\Delta t$ , the true answer is:

\int_0^t W_s\, dW_s = \frac{1}{2}W_t^2 - \frac{t}{2}

There is an extra $-t/2$ term that has no counterpart in ordinary calculus. This is not a rounding error or an approximation. It is the direct footprint of $(\Delta W)^2 \approx \Delta t$ the second-order term that refused to vanish. As you sum up more and more increments, the “extra” variance from squaring the Wiener increments accumulates into a deterministic correction of exactly $t/2$ .

We can verify this numerically:

import numpy as np
 
def ito_integral_simulation(T=1.0, N=10000, n_sims=5000):
    """
    Numerically estimates the Itô integral of W_s dW_s
    and compares it against the theoretical result: (1/2)*W_T^2 - T/2
    """
    dt = T / N
    results_ito  = []   # Numerical Itô integral
    results_theo = []   # Theoretical formula
 
    for _ in range(n_sims):
        dW = np.random.normal(0, np.sqrt(dt), N)
        W  = np.cumsum(dW)
        W  = np.concatenate([[0], W])   # W_0 = 0
 
        # Itô integral: use LEFT endpoint W[k] before increment dW[k]
        ito_sum = np.sum(W[:-1] * dW)
        results_ito.append(ito_sum)
 
        # Theoretical result
        results_theo.append(0.5 * W[-1]**2 - T / 2)
 
    print(f"Mean of Itô integral (numerical): {np.mean(results_ito):.4f}")
    print(f"Mean of theoretical formula:      {np.mean(results_theo):.4f}")
    print(f"Expected value (both should be ~0): {0:.4f}")
 
ito_integral_simulation()

Both outputs sit close to zero, confirming that the expected value of the Itô integral is zero. A key property that makes it useful for financial modelling. Across 5,000 simulated paths, the numerical left-endpoint sum and the theoretical formula $\frac{1}{2}W_T^2 - \frac{T}{2}$ agree, validating both the construction and the correction.

Itô’s Lemma: The New Chain Rule

Now we can fix the chain rule. Suppose we have a function $f(S, t)$ think of it as the price of an option that depends on the stock price $S$ and time $t$ . We want to know how $f$ changes as $S$ follows a GBM. In ordinary calculus, a Taylor expansion of $f$ to first order gives:

df = \frac{\partial f}{\partial S}\, dS + \frac{\partial f}{\partial t}\, dt

But $dS$ contains a $dW$ term. When we square it, we get $(dW)^2 \approx dt$ , a term we cannot ignore. Expanding fully and keeping all terms that survive in the limit gives us Itô’s Lemma:

df = \frac{\partial f}{\partial S}\, dS + \frac{\partial f}{\partial t}\, dt + \frac{1}{2} \sigma^2 S^2 \frac{\partial^2 f}{\partial S^2}\, dt

Or more compactly, after substituting in the GBM for $dS$ :

df = \left(\mu S \frac{\partial f}{\partial S} + \frac{\partial f}{\partial t} + \frac{1}{2}\sigma^2 S^2 \frac{\partial^2 f}{\partial S^2}\right) dt + \sigma S \frac{\partial f}{\partial S}\, dW_t

Breaking this down:

The $dt$ bracket contains the drift of $f$ : how it changes on average, per unit time.
The $dW_t$ term is the diffusion: the random component inherited from the stock price.
The extra $\frac{1}{2}\sigma^2 S^2 \frac{\partial^2 f}{\partial S^2}$ is the convexity correction — the critical innovation. It has no analogue in ordinary calculus, and it is entirely a consequence of $(\Delta W)^2 \neq 0$ .

Think of it this way: in a world of smooth, deterministic paths, second-order Taylor terms are negligible because they vanish faster than the first-order ones. But in a world where prices jump randomly those second-order fluctuations pile up into something that genuinely and measurably matters. Itô’s Lemma is the tool that accounts for them rigorously.

Solving the GBM: Applying Itô’s Lemma

Let us put Itô’s Lemma to work and actually solve the GBM for $S(t)$ . The trick is to pick the right function $f$ . Looking at the SDE $dS/S = \mu \, dt + \sigma \, dW$ , it looks like it should integrate to $\ln S$ . Let us verify what stochastic calculus actually says.

Set $f = \ln S$ . The required partial derivatives are:

\frac{\partial f}{\partial S} = \frac{1}{S}, \quad \frac{\partial f}{\partial t} = 0, \quad \frac{\partial^2 f}{\partial S^2} = -\frac{1}{S^2}

Plugging into Itô’s Lemma:

d(\ln S) = \frac{1}{S}\, dS + 0 + \frac{1}{2}\sigma^2 S^2 \cdot \left(-\frac{1}{S^2}\right) dt = \frac{dS}{S} - \frac{\sigma^2}{2}\, dt

Substituting $dS/S = \mu \, dt + \sigma \, dW$ :

d(\ln S) = \left(\mu - \frac{\sigma^2}{2}\right) dt + \sigma\, dW

This is just arithmetic Brownian motion in $\ln S$ — no randomness in the coefficient, no nonlinearity. We can integrate it directly over the interval $[0, t]$ :

\ln S(t) - \ln S(0) = \left(\mu - \frac{\sigma^2}{2}\right)t + \sigma W(t)

\boxed{S(t) = S_0 \exp\!\left[\left(\mu - \frac{\sigma^2}{2}\right)t + \sigma W(t)\right]}

This is the exact, closed-form solution to GBM. It tells us that the log of the stock price follows a Normal distribution which is why we say stock returns are log-normally distributed.

Notice the $-\sigma^2/2$ term, the Itô correction. Without it, the long-run drift of $\ln S$ would simply be $\mu$ . With it, the drift is $\mu - \sigma^2/2$ , which is always smaller than $\mu$ . This is not a bug; it is a deep truth about compounding under uncertainty. Large downward swings hurt more in percentage terms than equivalently large upward swings help, because losses are taken from a smaller base than gains. More volatility is, in a precise mathematical sense, a drag on long-run compounded growth.

import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import norm
 
# Parameters
S0 = 100
mu = 0.08
sigma = 0.20
T = 1.0
n_sims = 50000
 
# --- Simulate GBM paths using the exact Itô solution ---
np.random.seed(42)
W_T = np.random.normal(0, np.sqrt(T), n_sims)
 
# Exact closed-form solution via Itô's Lemma
S_T = S0 * np.exp((mu - 0.5 * sigma**2) * T + sigma * W_T)
 
# --- Plot distribution of final prices ---
fig, axes = plt.subplots(1, 2, figsize=(14, 5))
 
# Left: Distribution of S(T) — should be log-normal
axes[0].hist(S_T, bins=100, density=True, color='steelblue', alpha=0.7, edgecolor='none')
axes[0].axvline(np.mean(S_T), color='red', ls='--', label=f'Mean: ${np.mean(S_T):.2f}')
axes[0].axvline(np.median(S_T), color='orange', ls='--', label=f'Median: ${np.median(S_T):.2f}')
axes[0].set_title("Distribution of S(T): Log-Normal")
axes[0].set_xlabel("Stock Price at T=1 year ($)")
axes[0].set_ylabel("Density")
axes[0].legend()
 
# Right: Distribution of log returns — should be Normal
log_returns = np.log(S_T / S0)
axes[1].hist(log_returns, bins=100, density=True, color='darkorange', alpha=0.7, edgecolor='none')
 
# Overlay theoretical Normal distribution
x = np.linspace(log_returns.min(), log_returns.max(), 300)
theoretical_mean = (mu - 0.5 * sigma**2) * T
theoretical_std  = sigma * np.sqrt(T)
axes[1].plot(x, norm.pdf(x, theoretical_mean, theoretical_std),
             'k-', lw=2, label='Theoretical N(μ−σ²/2, σ²T)')
axes[1].set_title("Distribution of Log Returns: Normal")
axes[1].set_xlabel("Log Return ln(S(T)/S₀)")
axes[1].set_ylabel("Density")
axes[1].legend()
 
plt.tight_layout()
plt.show()
 
print(f"Theoretical mean of S(T): S0 * exp(μT) = ${S0 * np.exp(mu * T):.2f}")
print(f"Simulated  mean of S(T):  ${np.mean(S_T):.2f}")
print(f"\nDrift of log returns (Itô corrected):      {theoretical_mean:.4f}")
print(f"Drift of log returns (naive, without Itô): {mu * T:.4f}")

Three things to read from these results:

Left chart: The distribution of stock prices is visibly right-skewed. The mean ($108.33) sits to the right of the mode (~$102), which is the hallmark of a log-normal distribution. Prices cannot go below zero but can go arbitrarily high.
Right chart: The log returns sit perfectly on the Normal bell curve predicted by Itô’s Lemma. The curve is centred on 0.06, not 0.08 — the Itô correction is real and directly observable in simulation.
The output: The simulated mean of $108.21 matches the theoretical $108.33 closely. The 0.02 difference in the two drift estimates (0.06 vs 0.08) represents a 25% error in the growth rate that would accumulate meaningfully over long horizons if ignored.

Why Does It Matter?

Itô’s Lemma is not a mathematical curiosity. It is the direct engine behind one of the most important formulas in the history of finance: the Black-Scholes equation.

Fischer Black, Myron Scholes, and Robert Merton asked a deceptively simple question: if you hold an option (whose price is $f(S, t)$ ) and you continuously hedge it by holding some amount of the underlying stock $S$ , can you eliminate all the risk? By applying Itô’s Lemma to $f$ and carefully constructing a portfolio that cancels the $dW$ terms, they derived a Partial Differential Equation (PDE) for the option price. Solving that PDE gives the Black-Scholes formula, which is used to price trillions of dollars of derivatives every single day.

Every piece of that derivation rests on the logic we just built:

GBM models the stock price as a multiplicative, log-normal process.
Itô’s Lemma tells us how a function of that price (the option) evolves over time.
The $-\sigma^2/2$ correction is what makes compounding under uncertainty fundamentally different from compounding in a textbook.

The chain is now complete:

Coin flip → Binomial Tree → Wiener Process → Itô Integral → Itô’s Lemma → GBM closed-form solution → Black-Scholes

The next time you see an option price flicker on a trading screen, you are looking at the output of a stochastic differential equation, solved with a calculus built on the insight that randomness at the infinitesimal level is fundamentally different from anything Newton or Leibniz ever had to contend with. Every Greek: “Delta, Gamma, Vega, Theta”, is a partial derivative of the option price, calculated using exactly the framework we built in this post.

⚠️ Financial Education Disclaimer

The models and Python code in this post (including Geometric Brownian Motion and the Itô solution) are for educational and research purposes only.

Not Financial Advice: This content does not constitute professional financial or investment advice.

Model Limitations: GBM assumes constant volatility and log-normal returns. Real markets exhibit volatility clustering, fat tails, and sudden jumps that GBM does not capture. Models like Heston (stochastic volatility) and Merton (jump-diffusion) extend GBM to address these failures.

Risk Warning: Options and derivatives involve complex risks, including the potential for losses exceeding the initial investment. Always consult a qualified financial professional before trading.

Itô's Lemma Explained: Stochastic Calculus for Finance